open-mmlab
Something about the Sampler
Hi, I have read the code about the dist_samplers, and I have a question about it. It seems that in a epoch you don't sample all qualified images in a dataset. the number of images in a epoch seems depend on the 'num_instance' and the 'num of pid' in this dataset(For example, if we use market1501, we should sampler more than 10000 images in a epoch, but now we only sample 'num_instances * pids' images in a epoch). Will this have a bad effect on the model and final results due to the lack of images diversity in a epoch? 3Q >_<
A good question. Actually, I am not sure about it. I used such an identity sampler according to its wide usage in person re-id applications. You could try to modify it for more diverse samples for training. Welcome to show your results here.
